DrewJin0827/MBD-ComposedSFT

---
language:
- en
license: apache-2.0
task_categories:
- text-generation
tags:
- instruction-tuning
- supervised-fine-tuning
size_categories:
- 1M<n<10M
---

# MBD-ComposedSFT

## Dataset Summary

This dataset is a composed supervised fine-tuning (SFT) dataset that combines multiple high-quality instruction-following datasets. It comprises approximately 3,679,149 instruction tuning samples covering diverse domains including mathematics, coding, and general instruction following.

## Composition & Sources

The dataset is constructed from the following sources:

- smoltalk: 1,043,917 samples
- OpenMathInstruct 2 1M: 999,117 samples
- tulu 3 sft mixture: 937,730 samples
- opc sft stage2: 436,347 samples
- MathInstruct: 262,038 samples

## Data Format

Each record follows this structure:

```json
{
"query": "instruction or question text",
"response": "response or answer text"
}
```

## Citation

If you use this dataset, please cite the original source datasets:

- MathInstruct
- OpenMathInstruct-2
- Tulu 3
- SmolTalk
- OPC-SFT

## License

This dataset is released under the Apache 2.0 license.

## Acknowledgement

This dataset is derived from a thorough cleaning and re-processing of the ReFusion dataset. We acknowledge the creators and contributors of ReFusion for their foundational work, on which this dataset is based.